Agent Network Topologies

Why Topology Is an Architectural Decision

Before writing any agent code, the most important architectural question is: how will these agents connect? Topology determines throughput, fault tolerance, latency, and operational complexity. Getting it wrong means rewriting your entire message routing layer six months later.

Topology 1: Star (Hub-and-Spoke)

All agents communicate through a central orchestrator hub. No direct agent-to-agent communication.

flowchart TD
    HUB[Central Orchestrator Hub]
    A1[Specialist Agent 1]
    A2[Specialist Agent 2]
    A3[Specialist Agent 3]
    A4[Specialist Agent 4]
    A1 <-->|task + result| HUB
    A2 <-->|task + result| HUB
    A3 <-->|task + result| HUB
    A4 <-->|task + result| HUB
    style HUB fill:#EEF0F7,stroke:#6366F1,color:#0F172A

Best for: When a central orchestrator needs to maintain full visibility and control over all agents, customer support systems, approval workflows, systems where audit trails are required.

Tradeoffs:

• Hub is a single point of failure (mitigate with redundancy)
• Hub becomes throughput bottleneck at scale (mitigate with async queuing)
• Easy to monitor and debug, all traffic visible at one point

Topology 2: Mesh (Peer-to-Peer)

Every agent can communicate directly with every other agent. No central coordinator required.

flowchart LR
    A1[Risk Agent] <-->|findings| A2[Compliance Agent]
    A1 <-->|risk data| A3[Tax Agent]
    A2 <-->|compliance flags| A3
    A2 <-->|regulatory data| A4[Audit Agent]
    A3 <-->|tax implications| A4
    A1 <-->|exposure data| A4
    style A1 fill:#f0fdf9,stroke:#0D9488,color:#0F172A
    style A2 fill:#f0fdf9,stroke:#0D9488,color:#0F172A
    style A3 fill:#f0fdf9,stroke:#0D9488,color:#0F172A
    style A4 fill:#f0fdf9,stroke:#0D9488,color:#0F172A

Best for: Research networks, financial analysis systems, any domain where agents need to cross-reference each other’s findings without a natural hierarchy.

Tradeoffs:

• Maximum flexibility and no central bottleneck
• Complexity grows as O(N²), 10 agents = 90 potential connections
• Harder to debug (messages can take many paths)
• Requires robust message correlation (track which response answers which request)

Topology 3: Hierarchical (Tree)

Agents organized in tiers: executive → manager → worker. Each level delegates to the level below and reports upward.

flowchart TD
    CEO[CEO Agent\nstrategic decisions]
    VP1[VP Agent\nresearch division]
    VP2[VP Agent\nanalysis division]
    W1[Research Worker 1]
    W2[Research Worker 2]
    W3[Analysis Worker 1]
    W4[Analysis Worker 2]
    CEO --> VP1 & VP2
    VP1 --> W1 & W2
    VP2 --> W3 & W4
    W1 & W2 -->|results| VP1
    W3 & W4 -->|results| VP2
    VP1 & VP2 -->|summaries| CEO
    style CEO fill:#EEF0F7,stroke:#6366F1,color:#0F172A
    style VP1 fill:#EEF0F7,stroke:#818CF8,color:#0F172A
    style VP2 fill:#EEF0F7,stroke:#818CF8,color:#0F172A

Best for: Enterprise systems with clear organizational analogues, management consulting pipelines, content production at scale, systems where different tiers need different levels of context.

Tradeoffs:

• Natural for complex multi-phase work
• Information bottleneck at each tier, managers must summarize worker output, losing detail
• Latency compounds across tiers
• Excellent accountability, each node is responsible for its subtree

Topology 4: Pipeline (Sequential Stages)

Agents process work in a defined sequence. Each agent’s output is the next agent’s input.

flowchart LR
    IN([Raw Input]) --> A1[Ingest\nAgent]
    A1 --> A2[Transform\nAgent]
    A2 --> A3[Validate\nAgent]
    A3 --> A4[Enrich\nAgent]
    A4 --> A5[Publish\nAgent]
    A5 --> OUT([Output])
    style IN fill:#f0fdf9,stroke:#0D9488,color:#0F172A
    style OUT fill:#f0fdf9,stroke:#0D9488,color:#0F172A

Best for: ETL pipelines, document processing, content workflows (draft → review → approve → publish), any process that naturally has ordered stages.

Tradeoffs:

• Simple to reason about and debug
• Each stage can be scaled independently
• Sequential coupling: slow stage blocks everything downstream
• Easy to add stages without redesigning the network

Hybrid Topology: Real Enterprise Systems

Production systems combine topologies. A common pattern: hierarchical control plane + pipeline data flow + mesh for cross-functional collaboration.

flowchart TD
    subgraph Control["Hierarchical Control Plane"]
        ORCH[Orchestrator]
        M1[Manager: Data]
        M2[Manager: Analysis]
        ORCH --> M1 & M2
    end
    subgraph DataPipeline["Pipeline: Data Processing"]
        D1[Ingest] --> D2[Transform] --> D3[Validate]
    end
    subgraph MeshAnalysis["Mesh: Cross-functional Analysis"]
        AN1[Risk Analyst]
        AN2[Compliance Analyst]
        AN1 <-->|cross-reference| AN2
    end
    M1 --> DataPipeline
    M2 --> MeshAnalysis
    D3 -->|validated data| AN1 & AN2
    AN1 & AN2 -->|findings| ORCH
    style ORCH fill:#EEF0F7,stroke:#6366F1,color:#0F172A

Topology selection checklist:

• Clear sequential stages with dependencies → Pipeline
• Central control + audit requirements → Star
• Deep organizational hierarchy + large scale → Hierarchical
• Cross-functional collaboration without natural leader → Mesh
• Complex enterprise system → Hybrid (most common)

The topology you choose now shapes every implementation decision that follows. Choose based on your communication patterns, not on what’s easiest to implement first.